1 /*- 2 * SPDX-License-Identifier: Beerware 3 * 4 * ---------------------------------------------------------------------------- 5 * "THE BEER-WARE LICENSE" (Revision 42): 6 * <phk@FreeBSD.org> wrote this file. As long as you retain this notice you 7 * can do whatever you want with this stuff. If we meet some day, and you think 8 * this stuff is worth it, you can buy me a beer in return. Poul-Henning Kamp 9 * ---------------------------------------------------------------------------- 10 * 11 * $FreeBSD$ 12 */ 13 14 #ifndef _SYS_SMP_H_ 15 #define _SYS_SMP_H_ 16 17 #ifdef _KERNEL 18 19 #ifndef LOCORE 20 21 #include <sys/cpuset.h> 22 #include <sys/queue.h> 23 24 /* 25 * Types of nodes in the topological tree. 26 */ 27 typedef enum { 28 /* No node has this type; can be used in topo API calls. */ 29 TOPO_TYPE_DUMMY, 30 /* Processing unit aka computing unit aka logical CPU. */ 31 TOPO_TYPE_PU, 32 /* Physical subdivision of a package. */ 33 TOPO_TYPE_CORE, 34 /* CPU L1/L2/L3 cache. */ 35 TOPO_TYPE_CACHE, 36 /* Package aka chip, equivalent to socket. */ 37 TOPO_TYPE_PKG, 38 /* NUMA node. */ 39 TOPO_TYPE_NODE, 40 /* Other logical or physical grouping of PUs. */ 41 /* E.g. PUs on the same dye, or PUs sharing an FPU. */ 42 TOPO_TYPE_GROUP, 43 /* The whole system. */ 44 TOPO_TYPE_SYSTEM 45 } topo_node_type; 46 47 /* Hardware indenitifier of a topology component. */ 48 typedef unsigned int hwid_t; 49 /* Logical CPU idenitifier. */ 50 typedef int cpuid_t; 51 52 /* A node in the topology. */ 53 struct topo_node { 54 struct topo_node *parent; 55 TAILQ_HEAD(topo_children, topo_node) children; 56 TAILQ_ENTRY(topo_node) siblings; 57 cpuset_t cpuset; 58 topo_node_type type; 59 uintptr_t subtype; 60 hwid_t hwid; 61 cpuid_t id; 62 int nchildren; 63 int cpu_count; 64 }; 65 66 /* 67 * Scheduling topology of a NUMA or SMP system. 68 * 69 * The top level topology is an array of pointers to groups. Each group 70 * contains a bitmask of cpus in its group or subgroups. It may also 71 * contain a pointer to an array of child groups. 72 * 73 * The bitmasks at non leaf groups may be used by consumers who support 74 * a smaller depth than the hardware provides. 75 * 76 * The topology may be omitted by systems where all CPUs are equal. 77 */ 78 79 struct cpu_group { 80 struct cpu_group *cg_parent; /* Our parent group. */ 81 struct cpu_group *cg_child; /* Optional children groups. */ 82 cpuset_t cg_mask; /* Mask of cpus in this group. */ 83 int32_t cg_count; /* Count of cpus in this group. */ 84 int16_t cg_children; /* Number of children groups. */ 85 int8_t cg_level; /* Shared cache level. */ 86 int8_t cg_flags; /* Traversal modifiers. */ 87 }; 88 89 typedef struct cpu_group *cpu_group_t; 90 91 /* 92 * Defines common resources for CPUs in the group. The highest level 93 * resource should be used when multiple are shared. 94 */ 95 #define CG_SHARE_NONE 0 96 #define CG_SHARE_L1 1 97 #define CG_SHARE_L2 2 98 #define CG_SHARE_L3 3 99 100 #define MAX_CACHE_LEVELS CG_SHARE_L3 101 102 /* 103 * Behavior modifiers for load balancing and affinity. 104 */ 105 #define CG_FLAG_HTT 0x01 /* Schedule the alternate core last. */ 106 #define CG_FLAG_SMT 0x02 /* New age htt, less crippled. */ 107 #define CG_FLAG_THREAD (CG_FLAG_HTT | CG_FLAG_SMT) /* Any threading. */ 108 109 /* 110 * Convenience routines for building and traversing topologies. 111 */ 112 #ifdef SMP 113 void topo_init_node(struct topo_node *node); 114 void topo_init_root(struct topo_node *root); 115 struct topo_node * topo_add_node_by_hwid(struct topo_node *parent, int hwid, 116 topo_node_type type, uintptr_t subtype); 117 struct topo_node * topo_find_node_by_hwid(struct topo_node *parent, int hwid, 118 topo_node_type type, uintptr_t subtype); 119 void topo_promote_child(struct topo_node *child); 120 struct topo_node * topo_next_node(struct topo_node *top, 121 struct topo_node *node); 122 struct topo_node * topo_next_nonchild_node(struct topo_node *top, 123 struct topo_node *node); 124 void topo_set_pu_id(struct topo_node *node, cpuid_t id); 125 126 enum topo_level { 127 TOPO_LEVEL_PKG = 0, 128 /* 129 * Some systems have useful sub-package core organizations. On these, 130 * a package has one or more subgroups. Each subgroup contains one or 131 * more cache groups (cores that share a last level cache). 132 */ 133 TOPO_LEVEL_GROUP, 134 TOPO_LEVEL_CACHEGROUP, 135 TOPO_LEVEL_CORE, 136 TOPO_LEVEL_THREAD, 137 TOPO_LEVEL_COUNT /* Must be last */ 138 }; 139 struct topo_analysis { 140 int entities[TOPO_LEVEL_COUNT]; 141 }; 142 int topo_analyze(struct topo_node *topo_root, int all, 143 struct topo_analysis *results); 144 145 #define TOPO_FOREACH(i, root) \ 146 for (i = root; i != NULL; i = topo_next_node(root, i)) 147 148 struct cpu_group *smp_topo(void); 149 struct cpu_group *smp_topo_alloc(u_int count); 150 struct cpu_group *smp_topo_none(void); 151 struct cpu_group *smp_topo_1level(int l1share, int l1count, int l1flags); 152 struct cpu_group *smp_topo_2level(int l2share, int l2count, int l1share, 153 int l1count, int l1flags); 154 struct cpu_group *smp_topo_find(struct cpu_group *top, int cpu); 155 156 extern void (*cpustop_restartfunc)(void); 157 extern int smp_cpus; 158 /* The suspend/resume cpusets are x86 only, but minimize ifdefs. */ 159 extern volatile cpuset_t resuming_cpus; /* woken up cpus in suspend pen */ 160 extern volatile cpuset_t started_cpus; /* cpus to let out of stop pen */ 161 extern volatile cpuset_t stopped_cpus; /* cpus in stop pen */ 162 extern volatile cpuset_t suspended_cpus; /* cpus [near] sleeping in susp pen */ 163 extern volatile cpuset_t toresume_cpus; /* cpus to let out of suspend pen */ 164 extern cpuset_t hlt_cpus_mask; /* XXX 'mask' is detail in old impl */ 165 extern cpuset_t logical_cpus_mask; 166 #endif /* SMP */ 167 168 extern u_int mp_maxid; 169 extern int mp_maxcpus; 170 extern int mp_ncpus; 171 extern volatile int smp_started; 172 173 extern cpuset_t all_cpus; 174 extern cpuset_t cpuset_domain[MAXMEMDOM]; /* CPUs in each NUMA domain. */ 175 176 /* 177 * Macro allowing us to determine whether a CPU is absent at any given 178 * time, thus permitting us to configure sparse maps of cpuid-dependent 179 * (per-CPU) structures. 180 */ 181 #define CPU_ABSENT(x_cpu) (!CPU_ISSET(x_cpu, &all_cpus)) 182 183 /* 184 * Macros to iterate over non-absent CPUs. CPU_FOREACH() takes an 185 * integer iterator and iterates over the available set of CPUs. 186 * CPU_FIRST() returns the id of the first non-absent CPU. CPU_NEXT() 187 * returns the id of the next non-absent CPU. It will wrap back to 188 * CPU_FIRST() once the end of the list is reached. The iterators are 189 * currently implemented via inline functions. 190 */ 191 #define CPU_FOREACH(i) \ 192 for ((i) = 0; (i) <= mp_maxid; (i)++) \ 193 if (!CPU_ABSENT((i))) 194 195 static __inline int 196 cpu_first(void) 197 { 198 int i; 199 200 for (i = 0;; i++) 201 if (!CPU_ABSENT(i)) 202 return (i); 203 } 204 205 static __inline int 206 cpu_next(int i) 207 { 208 209 for (;;) { 210 i++; 211 if (i > mp_maxid) 212 i = 0; 213 if (!CPU_ABSENT(i)) 214 return (i); 215 } 216 } 217 218 #define CPU_FIRST() cpu_first() 219 #define CPU_NEXT(i) cpu_next((i)) 220 221 #ifdef SMP 222 /* 223 * Machine dependent functions used to initialize MP support. 224 * 225 * The cpu_mp_probe() should check to see if MP support is present and return 226 * zero if it is not or non-zero if it is. If MP support is present, then 227 * cpu_mp_start() will be called so that MP can be enabled. This function 228 * should do things such as startup secondary processors. It should also 229 * setup mp_ncpus, all_cpus, and smp_cpus. It should also ensure that 230 * smp_started is initialized at the appropriate time. 231 * Once cpu_mp_start() returns, machine independent MP startup code will be 232 * executed and a simple message will be output to the console. Finally, 233 * cpu_mp_announce() will be called so that machine dependent messages about 234 * the MP support may be output to the console if desired. 235 * 236 * The cpu_setmaxid() function is called very early during the boot process 237 * so that the MD code may set mp_maxid to provide an upper bound on CPU IDs 238 * that other subsystems may use. If a platform is not able to determine 239 * the exact maximum ID that early, then it may set mp_maxid to MAXCPU - 1. 240 */ 241 struct thread; 242 243 struct cpu_group *cpu_topo(void); 244 void cpu_mp_announce(void); 245 int cpu_mp_probe(void); 246 void cpu_mp_setmaxid(void); 247 void cpu_mp_start(void); 248 249 void forward_signal(struct thread *); 250 int restart_cpus(cpuset_t); 251 int stop_cpus(cpuset_t); 252 int stop_cpus_hard(cpuset_t); 253 #if defined(__amd64__) || defined(__i386__) 254 int suspend_cpus(cpuset_t); 255 int resume_cpus(cpuset_t); 256 #endif 257 258 void smp_rendezvous_action(void); 259 extern struct mtx smp_ipi_mtx; 260 261 #endif /* SMP */ 262 263 int quiesce_all_cpus(const char *, int); 264 int quiesce_cpus(cpuset_t, const char *, int); 265 void smp_no_rendezvous_barrier(void *); 266 void smp_rendezvous(void (*)(void *), 267 void (*)(void *), 268 void (*)(void *), 269 void *arg); 270 void smp_rendezvous_cpus(cpuset_t, 271 void (*)(void *), 272 void (*)(void *), 273 void (*)(void *), 274 void *arg); 275 #endif /* !LOCORE */ 276 #endif /* _KERNEL */ 277 #endif /* _SYS_SMP_H_ */ 278